Google, Blackstone Create TPU "as a Service" Business

Google and Blackstone’s TPU-as-a-service venture is important for any number of reasons:

• it turns TPUs from a mostly Google-hosted product into a broader external infrastructure platform

• strengthens Google’s push to monetize its custom silicon

• gives AI customers a non-Nvidia acceleration path

• might clarify the neocloud business model. 

Blackstone is committing $5 billion in equity and an initial 500 MW of capacity coming online in 2027. 

The move tends to ratify the GPU as a service market and provides an alternative to the Nvidia ecosystem, at least in the “bare metal” portion of the business. 

The venture also might intensify pricing pressure and reduce differentiation in the inference market. 

The venture also tests the durability of the neocloud business itself. Today, a global scarcity of high-end AI training and inference compute creates the basis for the market.

Neoclouds originally emerged as stopgaps to address the GPU shortage, but their bare-metal economics are fragile, being based on what most believe are temporary shortages of capacity. 

Perhaps their long-term viability hinges on their ability to move up the stack into AI-native services, which puts them in direct competition with hyperscalers. And some will note how little protection the business has, given the thin profit margins and high continuing capital investment. 

source: McKinsey 

Neoclouds have a strong demand story, but their business model is structurally difficult because they combine very high capital intensity with fast hardware depreciation and aggressive price competition. The result is a market that can grow fast while still being hard to make sustainably profitable.

The core problem is that graphics processing units are expensive, and their resale or rental value falls quickly as new generations arrive. 

McKinsey notes that over a typical five-year depreciation horizon, GPU-hour pricing can decline by half or more, which forces providers to recover capital quickly or risk stranded assets.

So neoclouds must keep raising capital to buy the next wave of chips even while the prior fleet is losing value. This makes cash flow, financing terms, and utilization rates far more important than simple revenue growth.

GPU clouds are not just chip businesses; they are power, cooling, networking, and operations businesses as well. High energy costs, high-density racks, and increasingly complex cooling requirements raise operating expense and add execution risk.

Up to this point, neoclouds are heavily dependent on Nvidia for the chips, networking ecosystem, and much of the software stack.

Google will test that thesis.

A big reason neoclouds emerged was that they could undercut hyperscalers on price and provisioning speed, sometimes by large margins. But hyperscalers are responding.

That means the initial “GPU scarcity arbitrage” is not a durable moat by itself.

The strategic tension is that investors often want neoclouds to move up the stack into managed services, orchestration, inference platforms, or sector-specific solutions. Those layers can improve retention and margins, but they also bring neoclouds into direct competition with hyperscalers that have deeper ecosystems and broader product bundles.

So the firms face a hard choice: stay close to bare-metal GPU rental, where margins are thin, or build higher-value services, where competition is tougher and sales cycles are longer.

That suggests a need to pioneer niche markets, such as sovereign compute and specialized workloads.

If Internet Was About "Disintermediation," AI "Remediates"

If disintermediation removing distributors in the value chain) was the hallmark impact of the internet on media and content, artificial intelligence arguably reinserts a new mediation layer in the value chain. 

If the internet tended to remove gatekeepers, AI is going to bring that function back.

The internet let creators, publishers, and businesses reach users directly.

• Authors could self-publish.

• Musicians could bypass record labels.

• Retailers could sell directly online.

• Newspapers could distribute without print trucks.

The dominant idea was that gatekeepers lost power and relevance.

AI seems to be inserting a new form of mediation between users and content, especially as agents proliferate:

• Users ask an AI instead of visiting websites

• Buyers rely on AI agents instead of browsing stores

• Students consult AI instead of textbooks or tutors

• Professionals ask AI instead of reading reports.

If the internet reduced friction between producer and consumer, allowing creators to reach consumers directly, AI inserts a new layer the AI agent). 

The practical consequences:

• Traffic shifts from websites to AI interfaces.

• Brands lose direct contact with customers.

• Original creators may receive less attribution.

• AI becomes the gatekeeper.

Still, both the internet and AI function in some similar ways:

• Reduce transaction costs

• Increase access

• Disrupt existing business models

• Favor large platforms. 

What is different is the way AI affects the monetization of content. Instead of search engine optimization, AI abstracts the sources of content.

Where SEO was intended to deliver relevant links, AI “gives you the answer.”

During the internet era, media companies lost control over distribution.

During the AI era, they may lose control over:

• Discovery

• Attribution

• Monetization

• Customer relationships

The risk is greater because AI can both create and summarize content, reducing the need to visit the original source.

But something greater than “distribution” or “discovery” is at work. Where the internet largely focused on “discovery” of content, AI agents will essentially “create” content.

As the internet forced content providers to change, so AI will force changes as well. Organizations must optimize not only for human customers, but also for AI systems.

Era	Core Trend	Central Dynamic
Internet	Disintermediation	Remove middlemen
Search	Aggregation	Organize information
Social Media	Platformization	Concentrate attention
AI	New intermediation	Machine agents become the middlemen

The shifts are from discovery to meaning; traffic volume to citation; site visits to inclusion in summary; active search to passive consumption. 

Feature	Internet Era Shift	AI Era Shift
Primary Value	Providing access to information.	Providing synthesis/meaning from information.
Core Metric	Clicks and Traffic.	Conversions and Authority/Citations.
Publisher Role	Destination (user visits your site).	Source (AI summarizes your work).
Advertising	Targeted banner/link ads.	Integrated conversational/agentic ads.
User Experience	Active searching and navigation.	Passive, conversational consumption.

The internet was about removing traditional gatekeepers. AI seems destined to insert a new layer into the value chain.

Participants will have to adapt. 

The primary challenge for publishers and content providers is “zero click.” When a user gets their answer directly on the search results page, they have no reason to visit the publisher’s site, leading to significant declines in organic traffic and ad impressions.

Advertisers now must design for agentic discovery as they once designed for search engines.

Platforms are performing a delicate balancing act: they need to keep users satisfied with quick, AI-driven answers while ensuring that content creators (the "web") don't disappear, as the AI needs that content to function.

Metric	Old Search Model	AI Search Era
Primary Goal	Traffic volume (clicks)	Brand authority & Citations
User Journey	Linear/Multi-click	Compressed/Conversational
Measurement	SEO rankings/CTR	Conversion-based/Brand impact
Publisher Value	Being a destination	Being a trusted source

One often hears lamentations about the unfairness of the new system, typically in the form of who gets the value from creating and distributing content. But nothing in the media and content businesses is guaranteed. Nobody has a “right” to audiences. 

New gatekeepers are coming, whether we like it or not.

How Do We Separate "Good" and "Bad" AI Implications?

What separates the "good" use of any technology and the "evil" use of that same technology? 

The simple answer is that most technology is morally inert. Human Intention is what separates the impact. 

But there is a sense in which “intention alone” is insufficient:

• Consequences matter independently of intent, which is why we have product liability laws

• At least some technologies are not entirely “neutral”

• landmines

• social media algorithms optimized for engagement

• Negligence (moral responsibility also extends to “what you should have foreseen”

• Externalities (climate, opioid addiction)

•  "Dual-use" (encryption; gain-of-function research) . 

So “intention is a “necessary but not sufficient” criteria for evaluating ethical implications. A fuller account could include:

• Design (What uses does the technology structurally enable or constrain?)

• Foreseeability (What harms were predictable?)

• Who benefits and who bears the risks?

• Systemic effects at scale.

So “intention” is the most important single factor in moral evaluation, but design, “affordances” (any property or feature of an object or environment that suggests and enables a specific action) and systemic effects generate moral responsibilities that exist independently of what anyone "meant."

Intent matters. But so do other consequences. The issue is how to create protections without weaponizing them (over-regulating; stifling; creating undue product liability laws).

Anxiety about artificial intelligence might be seen as “Luddite,” but AI concerns do include a legitimate focus on job markets, fairness, and safety. 

Job automation, economic inequality, bias, privacy, deepfakes, loss of human agency, concentration of power and longer-term potential risks are rational concerns.

On the other hand, there also is some admixture of resistance or skepticism about a new technology that might be shaped, but hardly seems possible to “stop.”

The signs are obvious, as objections are raised by:

• college students 

• younger workers

• community activists

• local residents

• political activists. 

The point is that "intent" is not a sufficient answer to the question of how we ascertain good and evil implications and uses of AI.

Despite what one intends, impact likely will have to be considered.

Ethical AI is Very Complicated

There are signs of anxiety about artificial intelligence that are well grounded but also “Luddite.” AI concerns do include a legitimate focus on job markets, fairness, and safety. 

Job automation, economic inequality, bias, privacy, deepfakes, loss of human agency, concentration of power and longer-term potential risks are rational concerns.

On the other hand, there also is some admixture of resistance or skepticism about a new technology that might be shaped, but hardly seems possible to “stop.”

The signs are obvious:

• college students 

• younger workers

• community activists

• local residents

• political activists. 

What separates the "good" use of any technology and the "evil" use of that same technology? 

The simple answer is that most technology is morally inert. Human Intention is what separates the impact. 

But there is a sense in which “intention alone” is insufficient:

• Consequences matter independently of intent, which is why we have product liability laws

• At least some technologies are not entirely “neutral”

• landmines

• social media algorithms optimized for engagement

• Negligence (moral responsibility also extends to “what you should have foreseen”

• Externalities (climate, opioid addiction)

•  "Dual-use" (encryption; gain-of-function research) . 

So “intention is a “necessary but not sufficient” criteria for evaluating ethical implications. A fuller account could include:

• Design (What uses does the technology structurally enable or constrain?)

• Foreseeability (What harms were predictable?)

• Who benefits and who bears the risks?

• Systemic effects at scale.

So “intention” is the most important single factor in moral evaluation, but design, “affordances” (any property or feature of an object or environment that suggests and enables a specific action) and systemic effects generate moral responsibilities that exist independently of what anyone "meant."

Intent matters. But so do other consequences. The issue is how to create protections without weaponizing them (over-regulating; stifling; creating undue product liability laws).

AI Bottlenecks Might be Shifting

The artificial intelligence buildout has reached the networking layer. Cisco's latest quarterly financial report provides an example.  

For three years, the capital expenditure story has been about GPUs. NVIDIA's data center revenue dominated the narrative. Hyperscalers committed hundreds of billions to compute clusters. But a GPU cluster without networking infrastructure is a warehouse full of processors that cannot talk to each other. Training runs that span tens of thousands of GPUs require switching fabrics that move data between them at speeds measured in terabits per second. Every additional GPU added to a cluster multiplies the networking demand, because the communication overhead scales faster than the compute scales linearly.

The AI infrastructure value chain is filling in a predictable order. Compute came first: NVIDIA's data center revenue grew from $15 billion in fiscal 2023 to nearly $200 billion in fiscal 2026, a roughly thirteenfold increase in three years. Networking is arriving now: Cisco's AI infrastructure run rate just quadrupled year over year. Storage will follow.

The bottlenecks are  moving down the stack from processors to the physical infrastructure that connects them. The backplane is where the bottleneck lives.

CAPE is an Issue, But How Much?

Nobody can know for certain--beyond the fact that U.S. financial markets are in historically above-average valuation levels--what could happen next. 

Some rationally expect a reversion to mean, which will mean lower valuations.

Others just as rationally argue that above-average valuations can persist for some time, and that a correction is not in store. AI might be among the reasons, if it changes growth expectations.

source:  Ark Invest

Consider the Cyclically Adjusted Price-to-Earnings Ratio (CAPE), widely considered a valuable long-term valuation metric. The current CAPE is high, suggesting caution and a likely correction to lower levels.  

But many analysts believe that changes in accounting rules since the early 2000s make the standard version look artificially high relative to its historical average. Adjusted, it might still be high, but not at internet bubble levels. 

The CAPE is calculated as the current S&P 500 Price divided by the average of 10 years of inflation-adjusted earnings.

The issue is that the denominator uses reported GAAP earnings, and those earnings have become more conservative over time, leading to a boost in CAPE that make comparisons with past levels misleading, the argument goes. 

Key accounting changes include: 

• Goodwill impairment rules (FAS 142, adopted in 2001)

• Large acquisition write-downs now hit earnings immediately.

• Before 2001, many such costs were spread over decades.

• This depresses modern earnings compared with earlier periods.

• Mark-to-market accounting

•  lk;juring crises, companies must recognize large non-cash losses.

• These can sharply reduce earnings even if long-term economics are less affected.

• One-time charges

• Restructuring costs and impairments are recognized more aggressively.

The result is that the denominator in today’s CAPE is lower than it would have been under earlier accounting rules, making the ratio appear higher.

Economist Jeremy Siegel argues for using National Income and Product Accounts instead of GAAP earnings, to better normalize over time. 

The standard CAPE can overstate market valuation materially because recent earnings include unusually large accounting write-downs by roughly 10 percent to 25 percent.

Others argue for using operating earnings rather than reported earnings, which also can adjust earnings by 15 percent.

Estimated Distortion	Standard CAPE 38	Adjusted CAPE
10%	38.0	34.2
15%	38.0	32.3
20%	38.0	30.4
25%	38.0	28.5

Using such methods, the market still appears expensive, but less so than it might appear. 

Other issues:

• Lower interest rates over long periods

• Higher profit margins

• Global diversification of large U.S. firms

• Greater use of stock buybacks instead of dividends

• Stronger institutional ownership and retirement savings flows.

These factors may justify a structurally higher "normal" CAPE than the 19th- and 20th-century average.

So some will argue a practical adjustment for accounting changes is to reduce the published Shiller P/E by 10 percent to 25 percent.

This suggests the market may still be richly valued, but not as dramatically overvalued as the unadjusted Shiller P/E implies.

It is a useful gauge of long-term valuation, but it is not a short-term market timing tool, as history shows that markets can continue to rise for years, even when the CAPE ratio is well above its historical average.

Several forces can keep markets rising despite expensive valuations:

• Earnings continue to grow

• Corporate profits may rise fast enough to justify higher prices

• Investor optimism and momentum

• Strong sentiment can sustain elevated valuations for extended periods

• Low interest rates

• When bond yields are low, investors are willing to pay more for equities

• New technologies can create expectations of stronger future growth

• Retirement contributions, buybacks, and institutional inflows can support prices.

Period	Approximate CAPE at Start	Years Until Major Peak	Additional Market Gain After CAPE Became Elevated	What Happened
1925–1929	25–32	4 years	+150% to +200%	Roaring Twenties speculation pushed valuations higher before the 1929 crash
1995–2000	25–44	5 years	+200% to +250%	Dot-com bubble drove extraordinary gains
2017–2021	30–38	4 years	+80% to +120%	Continued growth in Apple Inc., Microsoft Corporation, NVIDIA Corporation and other large-cap firms
2023–2026	Mid-30s (approx.)	Ongoing	Still developing	Strong enthusiasm around artificial intelligence and large technology firms

The point is that valuation is a poor short-term timing tool:

• A CAPE above average tells you expected long-term returns may be lower, but it does not predict when prices will stop rising

• Markets can stay expensive for years

• If profits rise rapidly, high valuations can become more sustainable

• Structural changes matter (lower inflation, global market reach, and dominant technology companies may justify higher valuation ranges than in earlier eras).

We still have to make our own choices about timing, though!